IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 4, NO.
2, APRIL 2017
Grid Integration of Wind Generation Considering
Remote Wind Farms: Hybrid Markovian and
Interval Unit Commitment
Bing Yan, Member, IEEE, Haipei Fan, Peter B. Luh, Fellow, IEEE, Khosrow Moslehi, Member, IEEE, Xiaoming
Feng, Member, IEEE, Chien Ning Yu, Member, IEEE, Mikhail A. Bragin, Member, IEEE, and Yaowen Yu
AbstractGrid integration of wind power is essential to reduce
fossil fuel usage but challenging in view of the intermittent
nature of wind. Recently, we developed a hybrid Markovian
and interval approach for the unit commitment and economic
dispatch problem where power generation of conventional units
is linked to local wind states to dampen the effects of wind uncer-
tainties. Also, to reduce complexity, extreme and expected states
are considered as interval modeling. Although this approach is
effective, the fact that major wind farms are often located in
remote locations and not accompanied by conventional units
leads to conservative results. Furthermore, weights of extreme
and expected states in the objective function are difficult to
tune, resulting in significant differences between optimization
and simulation costs. In this paper, each remote wind farm
is paired with a conventional unit to dampen the effects of
wind uncertainties without using expensive utility-scaled battery
storage, and extra constraints are innovatively established to
model pairing. Additionally, proper weights are derived through
a novel quadratic fit of cost functions. The problem is solved by
using a creative integration of our recent surrogate Lagrangian
relaxation and branch-and-cut. Results demonstrate modeling
accuracy, computational efficiency, and significant reduction of
conservativeness of the previous approach.
Index TermsBranch-and-cut, interval optimization, Markov
decision process, remote wind farms, surrogate Lagrangian
relaxation (SLR), unit commitment.
I. I NTRODUCTION
Manuscript received December 3, 2015; accepted May 25, 2016. This work
was supported in part by the Project Funded by ABB and U.S. National
Science Foundation (ECCS-1509666). Recommended by Associate Editor
Qinmin Yang.
Citation: B. Yan, H. P. Fan, P. B. Luh, K. Moslehi, X. M. Feng, C. N. Yu,
M. A. Bragin, and Y. W. Yu, Grid integration of wind generation considering
remote wind farms: hybrid markovian and interval unit commitment, IEEE/
CAA Journal of Automatica Sinica, vol. 4, no. 2, pp. 205215, Apr. 2017.
B. Yan, P. B. Luh, and M. A. Bragin are with the Department of
Electrical & Computer Engineering, University of Connecticut, Storrs,
CT 06269-4157, USA (e-mail: bing.yan@uconn.edu; peter.luh@uconn.edu;
mikhail.bragin@uconn.edu).
H. P. Fan was with the University of Connecticut, Storrs, CT 06269-4157,
USA (e-mail: haipei.fan@uconn.edu).
K. Moslehi and C. N. Yu are with ABB Inc., San Jose, CA 95134, USA
(e-mails: khosrow.moslehi@us.abb.com; Chien-Ning.Yu@us.abb.com).
X. M. Feng is with ABB Inc., Raleigh, NC 27606, USA (e-mail: xiaom-
ing.feng@us.abb.com).
Y. W. Yu is with ABB Enterprise Software, San Jose, CA 95134, USA
(e-mail: yaowen.yu@us.abb.com).
Any opinions, findings, and conclusions or recommendations expressed in
this material are those of the author(s) and do not necessarily reflect the views
of the National Science Foundation and ABB.
Digital Object Identifier 10.1109/JAS.2017.7510505
205
T O reduce greenhouse gas emissions and global warming
caused by fossil fuels, the use of renewable energy such
as wind is essential. In 2014 alone, more than 51 GW of wind
capacity was installed, bringing the global wind capacity to
nearly 370 GW, an increase of 14% since 2013 [1]. The U. S.
Department of Energys goal is to increase the nations wind
energy to 20% by 2030 [2]. A critical operation process of
wind integration is day-ahead unit commitment (UC). In the
UC process, independent system operators, which coordinate,
control and monitor power system operation within a single
or multiple states in the U. S., commit conventional units
with wind generation to meet the forecasted demand of the
following day. UC with high levels of wind integration is
challenging because of the intermittent nature of wind.
To incorporate wind uncertainties in UC, the literature
offers several approaches, e.g., stochastic programming, robust
optimization, interval optimization, and hybrid approaches. In
stochastic programming, uncertainties are modeled by rep-
resentative scenarios. However, it is difficult to select an
appropriate number of scenarios to balance modeling accuracy
and computational efficiency. While in robust optimization,
uncertainties are modeled by sets, and an optimal solution of
the worst-case realization is found, feasible for all realiza-
tions, leading to conservativeness. As for interval optimization,
uncertainties are modeled by intervals to capture bounds of
uncertain wind generation, where other realizations within
these bounds are guaranteed to be feasible. The approach
is computationally efficient, but results are still conservative.
There are also hybrid approaches of the above.
Recently, to solve the UC problem with uncertain wind
generation, we developed a hybrid approach [3]. The basic idea
is to divide power generation of conventional units into two
components: a Markovian one depending on local wind states
to dampen the effects of wind uncertainties, and an interval
one that manages extreme non-local states to capture constraint
bounds for solution feasibility. To reduce complexity, extreme
and expected states are considered as interval modeling. The
problem was solved by branch-and-cut. Results demonstrated
that the approach is effective, computationally efficient, and
less conservative as compared to other methods. However,
wind farms are often located in remote locations with high-
output wind resources, far from cities, where electricity de-
mand is high [4]. For example, in Texas, most wind farms
are in the windy western part, while load centers and most
conventional units are in the eastern part [5], [6]. With the
206
consideration of remote wind farms, this approach could
still be conservative since it degenerates to the pure interval
approach.
In this paper, our idea is to pair each remote wind farm with
a sufficiently large and not necessarily collocated conventional
unit with high ramp rates to dampen the effects of wind uncer-
tainties without using expensive utility-scaled battery storage.
However, with wind farms and units coupled, the problem
is difficult to formulate and to solve. Furthermore, weights
of extreme and expected states in the objective function are
difficult to tune, resulting in significant differences between
optimization and simulation costs.
Section II of this paper presents a literature review. Section
III establishes a novel formulation where each remote wind
farm is paired with a sufficiently large conventional unit with
high ramp rates. The pairing is based on heuristic rules,
not optimized, since the pairing itself is a combinatorial
problem. With non-local units treated as virtual local ones,
the constraints are established to formulate the pairing relation
and to model extreme generation. Section IV describes the
solution methodology, a creative integration of our recent
Surrogate Lagrangian Relaxation (SLR) and branch-and-cut
[7]. Additionally, proper weights are derived through a novel
quadratic fit of cost functions. Section V provides testing
results that include a simple system, and the IEEE 30-bus
and 118-bus systems. Numerical results demonstrate modeling
accuracy, computational efficiency, and significant reduction of
conservativeness of the previous approach. The formulation
established is general and can be applied to real-time UC as
well. This generic nature allows modeling of other intermittent
resources such as solar.
II. L ITERATURE R EVIEW
To solve UC with wind uncertainties, stochastic program-
ming, robust optimization, interval optimization, hybrid meth-
ods, and our earlier work are briefly reviewed in this section.
More detailed reviews of these methods are provided in [3].
A. Stochastic Programming
In stochastic programming, uncertainties are modeled by
representative scenarios [8][13]. To generate scenarios, wind
generation is typically assumed to follow a certain probability
distribution, and each scenario represents a sequenced real-
izations of wind uncertainties over the optimization horizon.
From the modeling point of view, the locations of wind farms
do not matter. In this method, optimization determines a
single set of UC decisions to satisfy all the selected scenarios
and multiple sets of dispatch decisions for the correspond-
ing scenarios. The objective is to minimize the sum of the
commitment cost and the expected dispatch cost with equal
weights. Decomposition methods, such as Benders decompo-
sition [10] and Lagrangian relaxation [11], are commonly used
to solve such problems. Since the number of scenarios in the
method can be extremely large even with discrete probability
distributions, scenario reduction is commonly used [14][16].
However, it is difficult to determine a proper number of
scenarios to balance modeling accuracy, solution feasibility,
and computational efficiency.
IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 4, NO. 2, APRIL 2017
B. Robust Optimization
In robust optimization, uncertainties are modeled by a pre-
determined set, where wind farms at different locations are
treated similarly [17][21]. Optimization determines a single
set of UC decisions to be feasible for all possible realizations
and a set of dispatch decisions against the worst-case realiza-
tion. Since the objective is to minimize the worst-case cost,
there is no need of weights. In robust optimization, Benders
decomposition is usually combined with other methods, such
as outer approximation [17] and cutting plane [18] to solve
the problems. This method gives conservative solutions, and
the models involve nonlinear min/max functions which require
much computational effort [17].
C. Interval Optimization
In interval optimization, uncertainties are modeled by closed
intervals in terms of upper and lower bounds [10], [22], where
the locations of wind farms are not a concern. In this method,
it requires a set of UC decisions to be feasible for all the
bounds, and two sets of dispatch decisions to be feasible for
the bounds of system demand and transmission capacity con-
straints captured by bounds of uncertain wind generation [22].
The objective is to minimize the sum of the costs associated
with the minimum and maximum wind generation under equal
weights. Since only the two sets of dispatch decisions and
bounds of system-wide constraints are considered, the method
requires less computational effort, and provides the lower and
upper bounds for the total cost. However, its optimal solution
is very sensitive to the uncertainty interval, which needs to be
carefully selected. A narrow interval may not cover the entire
uncertainty and the solution does not correspond to all possible
uncertain situations. A wide interval could lead to pessimistic
solutions where resources are not efficiently utilized.
D. Hybrid Methods
A hybrid stochastic and robust approach is developed in
[23], where dispatch decisions and constraints from both
stochastic programming and robust optimization are consid-
ered at the same time. By minimizing the weighted sum of
the costs from both approaches, this hybrid approach provides
more robust UC decisions than stochastic programming and
a lower simulation cost than robust optimization. System
operators can adjust the weights in the objective function
based on their preferences for the two methods. Its robust
optimization part is still nonlinear. A hybrid stochastic and
interval approach is developed in [24]. In this approach, in
the first few hours, stochastic programming is considered, and
in the remaining hours, interval optimization is used, where
their costs have equal weights. This hybrid approach provides
a lower simulation cost than either of the two methods,
while its interval optimization part remains conservative. For
neither stochastic/robust nor stochastic/interval approaches, the
locations of wind farms matter.
E. Our Previous Work
In our early work [25], without considering transmission
capacities, aggregated wind generation is molded as a Markov
YAN et al.: GRID INTEGRATION OF WIND GENERATION CONSIDERING REMOTE WIND FARMS: HYBRID MARKOVIAN AND
chain instead of scenarios to reduce complexity. However,
when considering transmission, wind generation at different
locations cannot be aggregated and has to be modeled as
a Markov chain at each node. To avoid explicitly consider-
ing all global states, interval optimization is synergistically
integrated with the Markovian approach in [3]. The basic
idea is to divide power generation of conventional units into
two components: a Markovian one depending on local wind
states to dampen the effects of wind uncertainties, and an
interval one that manages extreme non-local states to capture
constraint bounds for solution feasibility. In this way, the
Markovian component only depends on local wind states.
In the pure Markovian approach, wind generation depends
on all wind states, therefore the problem is combinatorial.
Consider a system with I wind farms, and assume that each
one has N states. For each local conventional unit, there
are N I dispatch decisions per hour where the complexity
increases exponentially with the number of wind farms. In
the hybrid approach, there are only N I dispatch decisions per
hour for each local conventional unit where the complexity
increases linearly with the number of wind farms, therefore the
computational complexity is dramatically reduced. In addition,
by making use of information provided by local states and their
transitions, this approach is less conservative than pure interval
optimization. The detailed complexity and conservativeness
comparison among the hybrid, pure Markovian and pure
interval approaches can be found in Section IV-B of [3].
With appropriate transformations, the problem is converted
to a linear form and solved by using branch-and-cut. In
Examples 1 and 2 of [3], a small system with three buses and
the IEEE 30-bus system are tested, respectively, and results
show that this approach is less conservative than pure interval
optimization. In Example 3, the IEEE 118-bus system is tested,
and the problem can be solved in about one minute, which
shows this method is computationally efficient. Also, the lower
relative difference between the simulation and optimization
costs demonstrates the modeling accuracy. However, with the
consideration of remote wind farms, the approach could still be
conservative since it degenerates to the pure interval approach.
In addition, weights of extreme and expected states in the
objective function are difficult to tune, resulting in significant
differences between optimization and simulation costs.
III. P ROBLEM F ORMULATION
Based on our early work [3], a novel formulation is
developed by pairing remote wind farms and conventional
units without using expensive utility-scaled battery storage
in this section. It contains three major sets of constraints at
unit, nodal, and system levels as presented in the first three
sections. The objective is to minimize the commitment cost
plus the weighted sum of the dispatch costs for the extreme
and expected states as shown in the last section.
A. Unit Level Constraints
Based on [3], consider a day-ahead energy market for an
independent system operator over 24 (T ) hours with each
hour indexed by t (1 t T ). At node i of the power
207
system, there are Ki ( 0) conventional units indexed by (i,
k) (0 k Ki ), and their properties such as cost functions
and capacities are assumed known. At some nodes, there
are also wind farms. The demand pL i (t) (MW) is assumed
known for node i at hour t. In the power system, there
are L transmission lines indexed by l (1 l L), where
line l has a transmission capacity flmax (MW). The UC
problem is to commit conventional units to meet the forecasted
demand of the following day while satisfying individual unit
and node constraints as well as transmission constraints. In
this section, constraints corresponding to UC decisions and
dispatch decisions at the unit level are formulated as follows.
1) Generation Level: Based on [3], the generation level of a
conventional uniti,k (dispatch decision) at time t is divided into
two components: Markovian generation pM i,k,ni (t) depending
on local wind state ni , and interval generation pIi,k,ni (t)
depending on extreme non-local states ni , i.e.,
pi,k,ni ,ni (t) = pM I
i,k,ni (t) + pi,k,ni (t) i; k; t. (1)
In the above, the minimum possible local state of node i is
denoted by min ni , and the maximum by max ni . Its minimum
non-local state is denoted as mi , which is a combination of
possible minimum states of other nodes. The maximum non-
local state is defined as Mi in a similar way. A global state
g at time t is a combination of wind generation states at all
nodes.
2) Startup Constraints: The binary startup variable ui,k (t)
equals 1 if and only if the unit is turned on from offline at
hour t, i.e.,
ui,k (t) xi,k (t) xi,k (t 1) i; k; t (2)
where xi,k (t) represents the on/off status (binary, 1 online
and 0 offline), and ui,k (t) represents the start-up status
(binary, 1 start-up and 0 otherwise).
3) Minimum up/down Time: The unit must remain online
or offline for its minimum up or down time, respectively. The
formulas in [27, eq. (3) and (5)] are used here.
4) Generation Capacity Constraints: If a unit is online, its
generation level should be within its minimum and maximum
values; otherwise, its generation level has to be zero, i.e.,
xi,k (t)pmin max
i,k pi,k,ni ,ni (t) xi,k (t)pi,k
i, k, t, ni i (t); ni {mi , Mi } (3)
where pmin max
i,k (MW) and pi,k (MW) are minimum and max-
imum generation levels, respectively, and i (t) is the set
of possible wind states of node i at hour t, i.e., i (t)
{ni |ni (t) > 0}, where ni (t) is the probability that wind
generation is at state ni during time t. The expression ni
i (t) and ni {mi , Mi } is omitted for the rest of the paper.
5) Ramp Rate Constraints: Ramp rate constraints require
that the change of the generation level cannot exceed the units
ramp rate between two consecutive hours. Based on [22, eq.
(21) and (22)], if the unit is online at hours t1 and t, then
for all possible state transitions and the two extreme non-local
208 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 4, NO. 2, APRIL 2017

states, the variation of generation levels cannot exceed its ramp
rate, i.e.,
pi,k,n0i ,n0i (t1) Ri,k pi,k,ni ,ni (t) pi,k,n0i ,n0i (t1)+Ri,k
(n0i , ni ) {(n0i , ni )|n0i (t), n0i ni > 0}
ni ; n0i ; i; k; t
(4)
where Ri,k (MW/ hour) denotes the ramp rate; n0i and n0i
denote the local and non-local states of node i at hour t1;
and denotes the transition probability from state n0i to state ni ,
which is established based on historical data and forecasted
weather. In addition, generation limits at start-up and shut-
down hours [28, eq. (11)] are merged with (4) based on [25].
The above constraints are for conventional units accompa-
nied with local wind farms. For the units paired with remote
wind farms, their Markovian generation depends on wind
states of the corresponding remote wind farms. For the units
not accompanied with local wind farms or paired with remote
ones, their generation only has the interval components.
The dispatch decisions of the expected state E (where wind
generation is at their expected values) will be also considered
in the objective function to be discussed later. With the same
set of commitment decisions and one set of dispatch decisions
pi,k,E (t), the constraints for the expected state can be easily
included as a set of deterministic constraints. These constraints
are not presented here for conciseness.
B. Nodal Level Constraints
Based on locations of wind farms and conventional units
in the system, there are five types of nodes which are associ-
ated with: wind farms and conventional units; remote wind
farms; conventional units paired with remote wind farms;
conventional units; and no wind farms or conventional units,
respectively. Let I LW , I RW , I LU , I N U and I N N denote the
sets of those nodes, respectively. The nodal level constraints
are presented below.
1) Load Shedding Constraints: When wind and conventional
generation cannot meet the system demand, the load has to be
shed, where the shed load cannot exceed the total load, i.e.,
LS,M LS,I
pLS L
i,ni ,ni (t) = pi,ni (t) + pi,ni (t) pi (t)
i; t; ni ; ni
where pLS,M LS,I LS
i,ni (t), pi,ni (t) and pi,ni ,ni (t) denote the Marko-
vian, interval and total load shedding of node i at time t,
respectively. The above constraints are for all nodes, while
there are no Markovian components for I N U and I N N .
2) Wind Curtailment Constraints: When wind and conven-
chain for simplicity. At each node, wind generation is dis-
cretized into N states, arranged in the ascending order.
3) Nodal Injections: The nodal injection for the node with
wind farms and conventional units equals wind generation
(after curtailment) plus conventional generation and minus the
demand (after load shedding), i.e.,
P
Pi,ni ,ni (t) = pW WC
i,ni (t) pi,ni (t) + pi,k,ni ,ni (t) + pLS
i,ni ,ni (t)
k
pL
i (t), i I LW ; t; ni ; ni .
(7)
For node i I RW , there is no conventional generation; for
node i I LU , there is no wind generation or curtailment; for
node i I N U , there are no wind related terms or Markovian
components; and for node i I N N , there are no wind or unit
related terms.
C. System Level Constraints
System level constraints consist of system demand and
transmission capacity constraints as follows.
1) System Demand Constraints: These constraints require
that the total wind generation plus the total conventional
generation equals the total system demand (after wind cur-
tailment and load shedding) at each hour, i.e., the sum of
nodal injections of all nodes, equals zero. Based on [22, eq.
(20)], the lower bound of the total wind generation happens
at the minimum global state m (when the outputs of all wind
farms are at their lower limits), while the upper bound happens
at the maximum global state M . As long as the minimum
and maximum states are satisfied, system demand is satisfied
for all other states. At the minimum state m, system demand
constraints are described as follows:
X
Pi,min ni ,mi (t) = 0 t. (8)
i
Similarly, the constraints at the maximum state M are:
X
Pi,max ni ,Mi (t) = 0 t. (9)
i
2) Transmission Capacity Constraints: These constraints
imply that the power flow through a line at each hour cannot
exceed its transmission capacity. In DC power flow, a line
(5) flow is a linear combination of nodal injections weighted by
generation shift factors (GSFs). Since GSFs could be positive
or negative, the nodal injection terms of (7) are regrouped to a
Markovian nodal injection consisting of those related to local
states:
P M
M
Pi,n i
(t) pW WC
i,ni (t) pi,ni (t) + pi,k,ni (t)
k

tional generation exceeds the system demand, wind generation + pLS,M L

i,ni (t) pi (t) i; t; ni (10)
has to be curtailed, where the curtailed wind power should be and an interval nodal injection related to non-local states:
less than the total, i.e., X
I
Pi,n (t) pIi,k,ni (t) + pLS,I
i,ni (t) i; t; ni . (11)
pW C W
i,ni (t) pi,ni , i I LW I RW ; t; ni (6) i
k

where pW C
i,ni (t)
and pW
i,ni (t)
denote the curtailed and total wind The above two equations are for nodes with wind farms
generation of node i under state ni at time t, respectively. and conventional units. For other types of nodes, their nodal
Based on [3], wind generation at different nodes is treated injections are regrouped with the corresponding components
separately and assumed to be modeled as independent Markov as discussed at the end of Section III-B.
YAN et al.: GRID INTEGRATION OF WIND GENERATION CONSIDERING REMOTE WIND FARMS: HYBRID MARKOVIAN AND 209

Bounds of Markovian flow levels are calculated based on wind farms and conventional units in (15) and (16) become
GSF signs and the corresponding extreme Markovian nodal linear.
injections, i.e., To linearize the min/max operations associated with other
P types of nodes, our idea is to treat non-local conventional units
[ail min Pi,n M
(t)]
iI LW I RW I LU :ail >0
ni i
which are paired with remote wind farms as virtual local units
P
+ [ail max Pi,n M
(t)] of the wind farms. In this way, the Markovian nodal injections
ni i
iI LW
P I RW I LU i
:al <0 of the nodes with remote wind farms and paired units have
[ail Pi,n
M
i
(t)] the following relationships, similar to (17):
iI LW I RW
P I LU

[ail max Pi,n M

i
(t)] M
Pi,n M
(t) + Pj,n M
(t) Pi,n M
(t) + Pj,n (t), i I RW
ni i 1 i 1
iI LW I RW I LU :ail >0 i i
P
+ [ail min Pi,nM
(t)] t. (12) j = IiPaired ; t; ni ; (ni 1) {ni 1|ni 1 (t) > 0}
i
iI LW I RW I LU :ail <0
ni (20)
where IiPaired denotes the node with a conventional unit that is
Bounds of interval flow levels can be obtained from the two paired with the remote wind farm at node i. Since Markovian
sets of interval nodal injections directly, i.e., nodal injections of the node with a remote wind farm and the
X
I
fl,m (t) = [ail Pi,m
I
(t)] l; t (13) node with the paired unit are coupled together, their extreme
i
Xi nodal injections may not happen at extreme wind states and
I
fl,M (t) = [ail Pi,M
I
i
(t)] l; t. (14) extra efforts are needed. With continuous decision variables
i pM,min
i, (t)/pM,max
i, (t) and binary ones xM,mini,ni (t)/xM,max
i,ni (t),
The above two interval flow levels are required to satisfy the the min/max operations can be linearized as follows,
two extreme Markovian flow levels in transmission capacity
PiM,max (t) Pi,n
M
(t)
constraints as formulated in (15) and (16), so that other states i

also satisfy transmission capacity constraints, i.e., PiM,max (t) Pi,n
M
i
(t) + 1 xM,max
i,ni (t) N
P P M,max
[ail min Pi,n
M
(t)] xi,ni (t) 1
ni i
iI LW I RW I LU :ail >0 ni
P PiM,min (t) Pi,n
M
(t)
+ [ail max Pi,n
M
i
I
(t)]+fl,g (t) i

ni
iI LW I RW I LU :ail <0
PiM,min (t) Pi,n
M
(t) 1 xM,min
i,n (t) N
flmax (t) l; t; g {m, M } P M,min i i
RW
P (15) xi,ni (t) 1, iI ; t; ni (21)
[ail max Pi,n
M
i
(t)] ni
ni
iI LW I RW I LU :ail >0
P where N is a large number [29].
+ [ail min Pi,n
M
i
I
(t)] + fl,g (t) Based on the above equations, for nodes with remote wind
ni
iI LW I RW I LU :ail <0
farms, the minimum (maximum) Markovian nodal injection
flmax (t) l; t; g {m, M }.
can be represented by pM,min (t) (pM,max (t)) at the optimum,
(16) i, i,
i.e.,
In the above, the min/max operations for the nodes with
wind farms and conventional units can be linearized based
M
min Pi,n i
(t) = PiM,min (t), i I RW ; t (22)
ni
on the monotonicity conjecture obtained in [3]. Consider two
possible local states at node i: state ni , and state ni 1. The
M
max Pi,n i
(t) = PiM,max (t), i I RW ; t (23)
ni
local state with lower wind generation provides a less or equal
Markovian nodal injection at the optimum, i.e., After including (20) and(21) as constraints and substituting
M M LW
the in/max operations with new variables as shown in (22)
Pi,n i 1
(t)
Pi,n (t),
iI and (23), the terms associated with the nodes with remote
i
(17)
t; ni ; (ni 1) {ni 1|ni 1 (t) > 0} wind farms in (15) and(16) become linear. For nodes with
where ni 1 (t) is the probability that wind generation is at conventional units paired with remote wind farms, their mini-
state ni 1 during time t. Based on the above conjecture, the mum (maximum) Markovian nodal injections can be modeled
minimum Markovian nodal injection happens at the minimum in a similar way. Since the linear constraints are equivalent
local wind state at the optimum, i.e., to min/max operations based on [29], the solution to the
linearized problem will be feasible to the original problem.
M M
min Pi,n i
(t) = Pi,min ni (t) i; t . (18)
ni

Similarly, the maximum Markovian nodal injection happens D. The Objective Function and Overall Problem
at the maximum local wind state at the optimum, i.e., The goal of the optimization problem is to minimize the
M
max Pi,n (t) = M
Pi,max ni (t) i; t . (19) commitment cost plus the expected dispatch cost. To avoid
i
ni
computational complexity, a weighted sum of the costs for
After including (17) as constraints and substituting the the extreme and expected states is used to approximate the
min/max operations with min/max states of nodal injections expected cost following the interval approach. The details of
as (18) and (19), the terms associated with the nodes with weight determination will be discussed in Section IV-B. The
210 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 4, NO. 2, APRIL 2017

)
resulting total cost to be minimized is the weighted dispatch P i M
I max
+ al min Pi,ni (t) +fl,g (t)fl (t)
cost plus the commitment cost, i.e., iI LW I RW I LU :ail <0
ni
P PP P PP
f (x) wm ni Cnmi (t) + wM ni CnMi (t) P P P E,N max
P i E

+ l,g (t) fl (t) al Pi (t)
P PPt Ei k P PP t i k
t l g
+wE Ci (t) + NL
(ui,k (t)Si,k + xi,k (t)Si,k ) i
t t
P P P E,P P i
i k i k
(24) + l,g (t) flmax (t) + al PiE (t)
t l g i
where (26)
Cnmi (t) = Ci,k pi,k,ni ,mi (t) + P LS pLS
i,ni ,mi (t) subject to (1)(7), (10), (11), (13), (14), (17)(25), and
minimum up/down time constraints. In the above, represents
CnMi (t) = Ci,k pi,k,ni ,Mi (t) + P LS pLS
i,ni ,Mi (t) multipliers relaxing system demand constraints, and repre-
LS LS,E sents multipliers relaxing transmission capacity constraints.
CiE (t) = Ci,k pE
i,k (t) + P pi (t). (25)
The relaxed problem is still inseparable into nodal subprob-
In the above, x represents all the decision variables (both
NL lems because of (20). To overcome this, the nodal subproblem
UC and dispatch decisions); Ci,k , Si,k and Si,k are gener-
with a remote wind farm (a conventional unit paired with a
ation, start-up and no-load costs, respectively; P LS is load
remote wind farm) will be solved with the decisions of other
shedding penalty; and m , M , and E are weights for
subproblems fixed as values obtained at the previous iteration.
dispatch costs of the minimum, maximum and expected states,
Now there are totally I subproblems (each one corresponding
respectively. The overall problem is to minimize the total cost
to one node) and will be solved by branch-and-cut [32], [33].
by selecting a single set of UC decisions and multiple sets of
By solving the relaxed problem, the dual function q be-
dispatch decisions of conventional units over 24 hours.
comes:
The above stochastic UC problem (1)(11), (13)(25),
q(, ) = min L(, , x). (27)
and minimum up/down time constraints is a mixed-integer x
linear optimization problem with binary decision variables Within SLR, instead of obtaining the dual value (27), a
{ui,k (t)} and {xi,k (t)}, and continuous decision variables surrogate dual value is obtained as follows,
LS,M LS,I
{pM I WC
i,k,ni (t)}, {pi,k,ni (t)}, {pi,ni (t)}, {pi,ni (t)}, {pi,ni (t)},
{pE LS,E W C,E L(k , k , xk ) = f (xk ) + k g(xk ) + k h(xk ). (28)
i,k (t)}, {pi (t)} and {pi (t)}. While wind farms
and conventional units are coupled through system demand In the above, xk is any feasible solution of the relaxed
constraints (8) and (9) and transmission capacity constraints problem at iteration k, and g(xk ) and h(xk ) are the surrogate
(15) and (16). subgradient vectors consisting of the following,
P k
IV. S OLUTION M ETHODOLOGY g m (xk ) = Pi,ni ,mi (t)
The above mixed-integer linear problem with coupling M k
Pi
k
g (x ) = Pi,n i ,Mi
(t)
constraints is solved by integration of our recent surrogate La- P i
grangian relaxation and branch-and-cut as presented in Section g E (xk ) = PiE k (t)
IV-A. To approximate the expected cost, proper weights are
i P i
hN k
g,l (x ) = fl
max
(t) M
al min Pi,n (t)
derived based on a simplified quadratic function in Section iI LW I RW I LU :ail >0
ni i

P i
IV-B. M
al max Pi,ni (t) fl,g I
(t)
n i
iI LW I RW I LU :ail <0
P
A. Surrogate Lagrangian Relaxation and Branch-and-cut [7], hP k max
ail max Pi,n
M
g,l (x ) = fl (t)+
ni i
(t)
[30], [31] iI LW I RW I LU :ail >0
i
P M I
The main idea of our surrogate Lagrangian relaxation + al min Pi,ni (t) + fl,g (t)
n i
iI LW I RW I LU :ail <0
approach is decomposition and coordination. After relaxing P i
system-wide coupling constraints, i.e., system demand (8) and hE,N
l (xk ) = flmax (t) al PiE (t)
Pi
(9) and transmission capacity constraints (15) and (16) by hE,P (xk ) = flmax (t) + ail PiE (t) .
l
introducing Lagrange multipliers, the relaxed problem is to i
minimize the following Lagrangian function L as, (29)
Since SLR does not require the relaxed problem to be
P m P
L(, , x) f (x) + (t) Pi,ni ,mi (t) fully optimized, surrogate subgradient directions may not form
t i acute angles with directions toward optimal multipliers, lead-
P M P P E P E
+ (t) Pi,ni ,Mi (t) + (t) Pi (t) ing to divergence. To guarantee that surrogate directions form
t i ( t i acute angles with directions toward the optimal multipliers,
PPP N P i M
the relaxed problem has to be sufficiently optimized, such
+ l,g (t) al min Pi,n (t)
t l g iI LW I RW I LU :ail >0
n i
i
that surrogate dual values (28) satisfy the following surrogate
)
P optimality condition:
M
ail max Pi,n I
(t) fl,g (t)flmax (t)
iI LW I RW I LU :ail <0
ni i
L(k , k , xk ) < L(k , k , xk1 ) (30)
(
PPP P P i M
where k and k are multipliers at the iteration k, and xk1
+ l,g (t) al max Pi,n (t)
t l g iI LW I RW I LU :ail >0
ni i
is a feasible solution at the iteration k1. Since the relaxed
YAN et al.: GRID INTEGRATION OF WIND GENERATION CONSIDERING REMOTE WIND FARMS: HYBRID MARKOVIAN AND 211

problem is not fully optimized and subgradient directions do determined by rewriting the unknown expected value fE (x),
not change much at each iteration, computational requirements the expected cost in the UC problem, by the three known
and zigzagging of multipliers are much reduced as compared values.
to traditional subgradient methods. For simplicity, it is assumed that wind generation follows a
In the method, multipliers are updated as uniform distribution, so does conventional generation x. Based
on the definition of f (x), fE (x) can be rewritten by using
k+1 = k + cSD,k g(xk ), k+1 = max 0, k +cT C,k h(xk ) . f (xmin ) and f (xmax ) as follows:
(31)
In the above, cSD,k and cT C,k are the two step sizes. fE (x)

It has been proven in [7] that the multipliers converge to a x2min + x2max + xmin xmax b(xmin + xmax )
= + +c
the optimum if the step sizes are updated as, 3 2
1 1
k1 SD,k1 k1 = ax2min + bxmin + c + ax2max + bxmax + c
cSD,k1
g(x ) c h(x ) 3 3
cSD,k=k , cT C,k=k



1 1
k
kg(x )k h(xk ) + axmin xmax + (bxmin + bxmax + 2c)
3 6
(32) 1 1 1
= f (xmin ) + f (xmax ) + axmin xmax
where 3 3 3
1 1 1
k = 1 , p=1 r + (bxmin + bxmax + 2c).
M kp k (33) 6
M 1; 0 < r < 1. (34)
The expected conventional generation xE can be described
The optimization stops when CPU time or the number of by xmin and xmax as follows:
iterations reach the pre-set stop time or the pre-set number.
Then, heuristics may be used to obtain feasible solutions xE = xmin + (1 )xmax
(35)
while a dual value provides a lower bound on the optimal xmin xE xmax ; 0 1.
cost. A duality gap can then be calculated by using the best
available feasible cost and the largest available dual value. For our UC problems, can be determined by system load,
Since the complexity of the new hybrid approach is similar and the minimum, maximum and expected wind generation
to the previous one, it will be more computationally efficient for each hour. Then by representing axmin xmax by f (xmin ),
than the pure Markovian approach based on the analysis in f (xE ), f (xmax ) as follows:
Section II. 1
axmin xmax= f (xE )2 f (xmin )(1)2 f (xmax )
2(1)
B. Weight Determination (1 ) (bxmin + bxmax + 2c)]
As mentioned earlier, the goal is to minimize the expected (36)
cost, i.e., commitment cost plus the expected dispatch cost. and substituting axmin xmax in (34) by (36), it can be obtained
To avoid computational complexity, a weighted sum of the that

costs for the minimum, maximum and expected states is used 1 1
as the optimization cost to approximate the expected cost. To fE (x) = f (xmin ) + f (xE )
3 6(1 ) 6(1 )
evaluate the optimal UC decisions, Monte Carlo simulation
runs are needed, and the simulation cost, the average cost 1 1
+ f (xmax ).
of all scenarios with UC decisions fixed at the optimal 3 6
solution, represents the expected cost. Modeling accuracy can (37)
be measured by the difference between the optimization and Therefore, the weights for f (xmin ), f (xE ), f (xmax ) to
simulation costs. Therefore it is crucial to determine proper approximate fE (x) are determined. In our UC problem, for
weights m , M , and E in (24), where the approximated simplicity, the average of 24 hours is used.
expected cost is obtained by optimization over 24 hours with
wind uncertainties. Our idea is to use a simplified quadratic V. N UMERICAL R ESULTS
function under a simple distribution and only one hour to The method presented above has been implemented by using
derive the weights, since the aggregated cost function of the optimization package IBM ILOG CPLEX Optimization
conventional units are approximately quadratic. Studio V 12.6.0.0 [33]. Testing has been performed on a PC
Consider an optimized cost function f (x) = ax2 + bx + with 2.90 GHz Intel Core (TM) i7 CPU and 16 GB RAM.
c (a > 0, x > 0) where conventional generation x follows Three examples are presented in this section. The first one
a certain distribution within a known range [xmin , xmax ]. is a small system to demonstrate the new approach is less
The expected conventional generation xE is also known based conservative than our previous approach (hybrid Markovian
on the initial wind states and state transaction matrices. The and interval), and to illustrate dispatch decisions. The second
values of f (xmin ), f (xmax ), and f (xE ) are assumed known as test example uses the IEEE 30-bus system to demonstrate com-
in the UC problem, representing the costs for the maximum, putational efficiency and modeling accuracy of our approach.
minimum, and expected wind states. Based on the generation And finally the IEEE 118-bus system is used to demonstrate
distribution and the cost function, proper weights can be scalability of our approach.
212 IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 4, NO. 2, APRIL 2017

A. Example 1: Small System weighted sum of the costs for the extreme and expected wind
Consider a 2-bus problem with one wind farm and one generation, is $1051. The total cost by using our new approach
conventional unit for one hour as shown in Fig. 1. The figure is $987. The results demonstrate that the new approach is less
also shows: 1) the generation values and probabilities of wind conservative than the previous approach. Dispatch decisions
generation states; 2) the minimum and maximum generation are also shown in Table I. Before pairing, Unit2,1 does not
levels and generation cost of the unit; 3) the load at the depend on wind states of the wind farm, so there is no
two nodes. Since only one hour is considered, time-coupling Markovian generation. After pairing, its Markovian generation
constraints such as ramp rate constraints and unit commitment depends on local state n1 of the wind farm at Node 1.
costs are ignored, and the time index is omitted. In our
B. Example 2: IEEE 30-bus System
previous hybrid Markovian and interval approach, the wind
farm and conventional unit are independent, while in our new In this example, the IEEE 30-bus system is used in a 24-
approach they are paired. Given xmin , xmax and xE , based on hour unit commitment based on the same parameters as in [3].
(37), the value of is 0.45, and the values of m , M , and It is assumed that there are 10 wind farms at Nodes 110,
E are 0.2, 0.13, and 0.67. To compare the conservativeness and 10 conventional units at Nodes 18 and 1112. The
of the two approaches, optimization results are provided in remote wind farms at Nodes 9 and 10 are paired with the
Table I. units at Nodes 11 and 12, respectively. As in [3], 10 states
Since the remote wind farm is paired with a conventional are used for individual wind farms, and their state transition
unit, the paired conventional unit works like expensive utility- matrices are established based on measured hourly generation
scaled battery storage to dampen the effects of wind uncer- data of 10 wind sites between April and September in 2006
tainties. With stochastic wind generation, the costs of extreme (non-winter season) selected from National Renewable Energy
states can be reduced through the pairing to dampen the wind Laboratorys Eastern Wind Dataset [34]. The distributions of
uncertain effects by using the new approach compared to the wind generation at steady states can be obtained based on
previous approach as shown in Table I. As for the expected the 10-state transition matrices. In the simulation, initial wind
state, the problem is deterministic, and the cost is not expected states are randomly generated based on these distributions.
to be reduced by the pairing. Generally, the costs with the With the expected state calculated based on 50-state transition
expected wind generation obtained by using our new approach matrices, the wind penetration level, calculated as the ratio of
and the previous approach may be not the same since the the total expected wind generation to the total demand without
solutions depend on wind states, properties of conventional wind curtailment and load shedding, is 5%.
units, weights in the objective function, etc. In this example, For SLR, the stopping criterion is the number of iterations,
the two costs with the expected wind generation happen to be e.g., 500 iterations. The total computational time includes
the same. By using the previous approach, the total cost, the data and model loading, subproblem solving, subgradient and
multiplier updating, and feasible solution searching time, while
solving time excludes the data and model loading time. For
branch-and-cut, the stopping criterion is to stop when the total
time or the MIP gap reaches its respective preset limit.
To evaluate optimal UC decisions, 1000 Monte Carlo sim-
ulation runs are performed with 50-state transition matrices.
Modeling accuracy is measured by the absolute percentage
error (APE), the ratio of the absolute difference between
optimization and simulation costs to the simulation cost. The
standard deviation (STD) of scenario costs reflects its varia-
tion. Results with a 5% wind penetration level are summarized
in Table II (m , M , and E are 0.05, 0.05, and 0.9).
Fig. 1. The two-bus transmission network for Example 1.
Compared with our previous approach, the feasible cost of
TABLE I the new one is reduced by 0.8% with less wind curtailment,
O PTIMIZATION R ESULTS FOR E XAMPLE 1 indicating that our approach is less conservative. By using
SLR and branch-and cut, a near-optimal solution with a 1.1%
Approach New Previous
duality gap is obtained in about 505 seconds. A solution
Cost ($) CostTotal 987.15 1,050.9 with a much higher mixed-integer programming (MIP) gap
Costm 1080 1350
is found after 1200 seconds by pure branch-and-cut, thereby
CostE 1020 1020
indicating that our approach is more computationally efficient.
CostM 675 750
In addition, our approach is also accurate, as its absolute
Dispatch decisions (MW) pM 30 / percentage error is about 0.3%.
2,1,1
pM2,2,3 20 /
pM 25 /
C. Example 3: IEEE 118-bus System
2,1,3
pI2,1,mi 15 45 In this example, the IEEE 118-bus system [35] with 54
pI2,1,Mi 0 25 conventional units is tested. It is assumed that there are 10
pE2,1 34 34 wind farms, 6 of which are accompanied with conventional
YAN et al.: GRID INTEGRATION OF WIND GENERATION CONSIDERING REMOTE WIND FARMS: HYBRID MARKOVIAN AND 213

TABLE II
R ESULTS FOR E XAMPLE 2: IEEE 30- BUS W ITH A 5% W IND P ENETRATION L EVEL
Approach New Previous
SLR+ B & C Pure B & C Pure B & C
Optimization Cost (k$) 305.728 460.296 308.149
Penalty (k$) ($5000/MWh) 0 / 0.14
Curtailed wind (MWh) 8.11 / 15.93
Total time (s) 505 1200 122
Solving time (s) 283.6 / /
Lower bound (k$) 302.634 286.710 307.906
Gap (%) 1.01 37.71 0.08

UC cost (k$) 95.41 / 97.66

Simulation Cost (k$) 306.715 309.176

APE 0.32 0.33
STD (k$) 1.51 1.5
Penalty (k$) ($5000/MWh) 0 0
Curtailed wind (MWh) 0 0

TABLE III
R ESULTS FOR E XAMPLE 3: IEEE 118- BUS
Wind penetration level 5% 15% 25%
Optimization (New by SLR) Cost (k$) 929.425 809.119 707.601
Penalty (k$) ($5000/MWh) 0.036 0.025 0
UC cost (k$) 12.56 12.50 11.92
Curtailed wind (MWh) 38.58 1583.28 4361.33
Total time (s) 935 1195 1525
Solving time (s) 786 1054 1400
Lower bound (k$) 927.347 802.874 699.407
Gap (%) 0.22 0.77 1.15

Optimization (Previous by B&C) Cost by (k$) 936 936 815 931 711 604
Reduced by the new approach (%) 0.8 0.83 0.63
Total time (s) 1500 2000 2000
Gap by (%) N/A 19 29.7

Simulation (New) Cost (k$) 926.988 808.601 705.068

APE 0.26 0.064 0.36
STD (k$) 12.64 37.80 51.02
Penalty (k$) ($5000/MWh) 0 0 0
Curtailed wind (MWh) 0 4.71 776.62

units and 4 are at remote locations. Each of the remote
wind farms is paired with a sufficiently large unit with
high ramp rates out of the 54 units. Wind state transition
matrices are obtained similarly as in Example 2. As in [3], the
quadratic cost curves of conventional units are approximated
by piecewise linear cost curves with three blocks. The hourly
system demand values in percent of peak system demand
are calculated based on corresponding factors for summer
weekdays of the IEEE Reliability Test System [36]. The
stopping iteration number is 1300. The results with three levels
of wind penetration, i.e., 5%, 15% and 25%, are summarized
in Table III as follows m , M , and E are: 0.05, 0.05, and
0.9; 0.05, 0.15, and 0.8; and 0.06, 0.24, and 0.7, respectively).
As compared with our previous approach, the total costs
obtained by the new approach are reduced by 0.6% to 0.8% for
each level of wind penetration. For practical power systems,
such as the market for Midcontinent Independent System
Operator, the total cost of a typical winter day is $42 m to
$63 m, while $52 m to $73 m for a typical summer day [37].
In addition, absolute percentage errors are all within 0.5%,
demonstrating the modeling accuracy. Compared with branch-
and-cut, our approach can obtain near-optimal solutions in
acceptable amount of time, demonstrating that our approach is
computationally efficient. It is also shown that the higher the
level of wind penetration, the longer the computational time.
For the IEEE 118-bus system with a 5% wind penetration
level, the total computational time is about 935 seconds, while
it is 505 seconds for the IEEE 30-bus system in Example 2.
With the same level of wind penetration, the computational
time is nearly linear to the number of wind farms and
conventional units. Therefore our new approach is scalable
for large systems.
VI. C ONCLUSION
This paper develops a new hybrid Markovian and interval
approach to solve the UC problem with the consideration of
remote wind farms. Since major wind sites are often located
in remote locations without accompanying conventional units,
214
our idea is to pair each remote wind farm with a sufficiently
large unit with high ramp rates to dampen the effects of wind
uncertainties without using expensive utility-scaled battery
storage. Extra constraints are established for pairing and proper
weights are derived through a simple quadratic function. The
problem is solved by integration of our recent surrogate La-
grangian Relaxation and branch-and-cut for near-optimal so-
lutions. Numerical results demonstrate that the new approach
is effective in terms of solution feasibility, modeling accuracy,
and computational efficiency. The formulation established is
general and can model real-time UC as well. Its generic nature
allows for modeling of wind and other intermittent resources
such as solar.
R EFERENCES
[1] Global wind report C Annual market update 2014, Global Wind
Energy Council, Brussels, Belgium, March 2015. [Online]. Available:
http://www.gwec.net/wp-content/uploads/2015/03/GWEC Global Wind
2014 Report LR.pdf
[2] U.S. Department of Energy. (2008, July). 20% wind energy
by 2030: Increasing wind energys contribution to U.S.
electricity supply. DOE/GO-102008-2567. [Online]. Available:
http://www.nrel.gov/docs/fy08osti/41869.pdf
[3] Y. W. Yu, P. B. Luh, E. Litvinov, T. X. Zheng, J. Y. Zhao, and F. Zhao,
Grid integration of distributed wind generation: Hybrid Markovian
and interval unit commitment, IEEE Trans. Smart Grid, vol. 6, no. 6,
pp. 30613072, Nov. 2015.
[4] H. J. Wagner and J. Mathur, Introduction to wind energy systems: basics,
technology and operation, Springer Science & Business Media, second
edition, 2012, pp.3.
[5] M. D. Franco, Nearly Completed CREZ Lines Unlock Wind Conges-
tion, North American Wind Power, Magazine, Vol. 10, No.6, 2013. [On-
line]. Available: http://www.nawindpower.com/issues/NAW1307/FEAT
01 Nearly Completed CREZ Lines Unlock Wind Congestion.html
[6] J. Muyskens, D. Keating, and S. Granados. (2015,
July). Mapping how the United States generates its
electricity. The Washington Post. [Online]. Available at
https://www.washingtonpost.com/graphics/national/power-plants/
[7] M. A. Bragin, P. B. Luh, J. H. Yan, N. P. Yu, and G. A. Stern,
Convergence of the surrogate Lagrangian relaxation method, J.
Optimizat. Theory Appl., vol. 164, no. 1, pp. 173201, Jan. 2015.
[8] F. Bouffard and F. D. Galiana, Stochastic security for operations
planning with significant wind power generation, IEEE Trans. Power
Syst., vol. 23, no. 2, pp. 306316, May 2008.
[9] P. A. Ruiz, C. R. Philbrick, E. Zak, K. W. Cheung, and P. W. Sauer,
Uncertainty management in the unit commitment problem, IEEE
Trans. Power Syst., vol. 24, no. 2, pp. 642651, May 2009.
[10] L. Wu, M. Shahidehpour, and Z. Y. Li, Comparison of scenario-based
and interval optimization approaches to stochastic SCUC, IEEE Trans.
Power Syst., vol. 27, no. 2, pp. 913921, May 2012.
[11] A. Papavasiliou, S. S. Oren, and R. P. ONeill, Reserve requirements
for wind power integration: A scenario-based stochastic programming
framework, IEEE Trans. Power Syst., vol. 26, no. 4, pp. 21972206,
Nov. 2011.
[12] M. S. Li, Q. H. Wu, T. Y. Ji, and H. Rao, Stochastic multi-objective
optimization for economic-emission dispatch with uncertain wind power
and distributed loads, Electr. Power Syst. Res., vol. 116, pp. 367373,
Nov. 2014.
[13] E. M. Constantinescu, V. M. Zavala, M. Rocklin, S. Lee, and M.
Anitescu, A computational framework for uncertainty quantification
and stochastic optimization in unit commitment with wind power gener-
ation, IEEE Trans. Power Syst., vol. 26, no. 1, pp. 431441, Feb. 2011.
IEEE/CAA JOURNAL OF AUTOMATICA SINICA, VOL. 4, NO. 2, APRIL 2017
[14] J. Dupaca, N. Growe-Kuska, and W. Romisch, Scenario reduction in
stochastic programming, Mathem. Progr., vol. 95, no. 3, pp. 493511,
Mar. 2003.
[15] N. Growe-Kuska, H. Heitsch, and W. Romisch, Scenario reduction and
scenario tree construction for power management problems, in Proc.
2013 IEEE Bologna Power Tech Conf., Bologna, Italy, 2003.
[16] J. M. Morales, S. Pineda, A. J. Conejo, and M. Carrion, Scenario
reduction for futures market trading in electricity markets, IEEE Trans.
Power Syst., vol. 24, no. 2, pp. 878888, May 2009.
[17] D. Bertsimas, E. Litvinov, X. A. Sun, J. Y. Zhao, and T. X. Zheng,
Adaptive robust optimization for the security constrained unit commit-
ment problem, IEEE Trans. Power Syst., vol. 28, no. 1, pp. 5263,
Feb. 2013.
[18] L. Zhao and B. Zeng, Robust unit commitment problem with demand
response and wind energy, in Proc. 2012 IEEE Power and Energy
Society General Meeting, San Diego, California, 2012, pp. 18.
[19] R. W. Jiang, J. H. Wang, and Y. P. Guan, Robust unit commitment
with wind power and pumped storage hydro, IEEE Trans. Power Syst.,
vol. 27, no. 2, pp.800810, May 2012.
[20] C. Lee, C. Liu, S. Mehrotr, and M. Shahidehpour, Modeling trans-
mission line constraints in two-stage robust unit commitment problem,
IEEE Trans. Power Syst., vol. 29, no. 3, pp. 12211231, May 2014.
[21] R. W. Jiang, J. H. Wang, M. H. Zhang, and Y. P. Guan, Two-stage
minimax regret robust unit commitment, IEEE Trans. Power Syst.,
vol. 28, no. 3, pp. 22712282, Aug. 2013.
[22] Y. Wang, Q. Xia, and C. Q. Kang, Unit commitment with volatile
node injections by using interval optimization, IEEE Trans. Power
Syst., vol. 26, no. 3, pp. 17051713, Aug. 2011.
[23] C. Y. Zhao and Y. P. Guan, Unified stochastic and robust unit com-
mitment, IEEE Trans. Power Syst., vol. 28, no. 3, pp. 33533361,
Aug. 2013.
[24] Y. Dvorkin, H. Pandzic, M. A. Ortega-Vazquez, and D. S. Kirschen,
A hybrid stochastic/interval approach to transmission-constrained unit
commitment, IEEE Trans. Power Syst., vol. 30, no. 2, pp. 621631,
Mar. 2015.
[25] P. B. Luh, Y. W. Yu, B. J. Zhang, E. Litvinov, T. X. Zheng, F. Zhao,
J. Y. Zhao, and C. C. Wang, Grid integration of intermittent wind
generation: a Markovian approach, IEEE Trans. Smart Grid, vol. 5,
no. 2, pp. 732741, Mar. 2014.
[26] P. Luh, H. P. Fan, K. Moslehi, X. M. Feng, M. Bragin, Y. W. Yu, C. N.
Yu, and A. Mousavi, An Extended Hybrid Markovian and Interval Unit
Commitment Considering Renewable Generation Uncertainties, U. S.
Federal Energy Regulatory Commission, Washington, DC, Jun. 2015.
[27] D. Rajan and S. Takriti, Minimum up/down polytopes of the unit
commitment problem with start-up costs, IBM Research Report, 2005.
[28] X. Guan, P. B. Luh, H. Yan, and J. A. Amalfi, An optimization-based
method for unit commitment, Int. J. Electr. Power Energy Syst., vol. 14,
no. 1, pp. 917, Feb. 1992.
[29] FICO. (2009, June). MIP formulations and linearizations. [Online].
Available: http://www.fico.com/en/node/8140?file=5125
[30] M. A. Bragin, P. B. Luh, J. H. Yan, and G. A. Stern, Surrogate
Lagrangian relaxation and branch-and-cut for unit commitment with
combined cycle units, in Proc. IEEE Power and Energy Society
General Meeting, National Harbor, Maryland, USA, 2014, pp. 15.
[31] M. A. Bragin, P. B. Luh, J. H. Yan, and G. A. Stern, Novel exploita-
tion of convex hull invariance for solving unit commitment by using
surrogate Lagrangian relaxation and branch-and-cut, in Proc. IEEE
Power and Energy Society General Meeting, Denver, Colorado, USA
2015, pp. 15.
[32] R. E. Bixby, M. Fenelon, Z. H. Gu, E. Rothberg, and R. Wunderling,
MIP: Theory and practice closing the gap, in System Modelling
YAN et al.: GRID INTEGRATION OF WIND GENERATION CONSIDERING REMOTE WIND FARMS: HYBRID MARKOVIAN AND
and Optimization, M. J. D. Powell and S. Scholtes, Eds. US: Springer,
2000, pp. 1949.
[33] IBM ILOG CPLEX Optimization Studio Information Center, IBM
ILOG, 2013. [Online]. Available: http://www.ibm.com/support/knowled
gecenter/SSSA5P 12.5.1/maps/ic-homepage.html
[34] The National Renewable Energy Laboratory, Eastern Wind Dataset,
2010. [Online]. Available: http://www.nrel.gov/electricity/transmission/
eastern wind methodology.html
[35] IEEE 118-bus system. [Online]. Available: http://motor.ece.iit.edu/data/
[36] C. Grigg, P. Wong, P. Albrecht, R. Allan, M. Bhavaraju, R. Billinton,
Q. Chen, C. Fong, S. Haddad, S. Kuruganty, W. Li, R. Mukerji, D.
Patton, N. Rau, D. Reppen, A. Schneider, M. Shahidehpour, and C.
Singh, The IEEE reliability test system-1996. A report prepared by the
reliability test system task force of the application of probability methods
subcommittee, IEEE Trans. Power Syst., vol. 14, no. 3, pp. 10101020,
Aug. 1999.
[37] MISO Energy, Day-Ahead Pricing Report, Report, Jan 1117 and July
1218, 2015, [Online]. Available: https://www.misoenergy.org/Library/
MarketReports/Pages/MarketReports.aspx
Bing Yan (S11-M17) received the B.S. degree
from Renmin University of China in 2010, M.S.
and Ph.D. degrees from University of Connecticut
in 2012 and 2016, respectively. She is currently an
assistant research professor in the Department of
Electrical and Computer Engineering, University of
Connecticut.
Her research interests include power system opti-
mization, grid integration of renewables (wind and
solar), energy-based operation optimization of dis-
tributed energy systems, and scheduling of manu-
facturing systems. Corresponding author of this paper.
Haipei Fan received her B.E. degree from East
China University of Science and Technology in
2013, and M.S. from University of Missouri-
Columbia in 2014.
Her current research interests include power sys-
tem optimization and grid integration of renewable
energy.
Peter B. Luh (S77M80SM91F95) received
his B.S. degree from National Taiwan University,
M.S. degree from M.I.T., and Ph.D. degree from
Harvard University. He has been with the University
of Connecticut since 1980, and is the SNET profes-
sor of communications & information technologies.
His interests include smart power systems smart
grid, design of auction methods for electricity mar-
kets, effective renewable (wind and solar) integration
to the grid, electricity load and price forecasting with
demand response, and micro grid. He is a fellow of
IEEE, was the vice president of publication activities for the IEEE Robotics
and Automation Society.
215
Khosrow Moslehi received his Ph.D. degree from
the University of California, Berkeley. He is the
director of product management at ABB Network
Control in Santa Clara, CA. Dr. Moslehi has 30 years
of experience in R&D.
His research interests include power system anal-
ysis and optimization, system integration and archi-
tecture, electricity markets, and the smart grid.
Xiaoming Feng (M91) received his B.S. degree
from Xian Jiaotong University in 1982, and he
obtained his M.S. and Ph.D. degree in electrical &
computer engineering from Ohio University in 1986
and 1989, respectively. He is a principal consulting
R&D engineer with ABBs research lab Electrical
System Technology Institute, Raleigh, NC. He has
more than 20 years of industry experience working
as management consultant, R&D engineer, software
developer, and software product manager. In recent
years, he has been engaged in consulting and re-
search in areas ranging from deregulated energy market analysis, using genetic
methodology in transmission and distribution system planning, power quality,
and distributed power and renewable energy.
His research interests include simulation, analysis, planning, and optimiza-
tion of electric power transmission and delivery systems using advanced
simulation, optimization, and probabilistic techniques.
Chien Ning Yu received his B.S. degree from
National Taiwan University in 1992, and he obtained
his M.S. and Ph.D. degree in mechanical engineering
from MIT in 1996 and 1999 respectively. He is
currently a senior consulting engineer in ABB Inc.,
Santa Clara, CA. He has been actively involved
in design and implementation of Texas ERCOT,
Ontario IESO, South Korea KPX, East China ECG,
Philippine WESM, and Western Australia IMO mar-
ket applications. His research interests include power
systems operation and economy, dynamics of com-
plex networks and automatic control theories.
Mikhail A. Bragin (S11-M17) received his B.S.
and M.S. degrees in mathematics from the Voronezh
State University, Russia, in 2004, the M.S. degree
in physics and astronomy from the University of
Nebraska-Lincoln, USA, in 2006, and the M.S. and
Ph.D. degree in electrical and computer engineering
from the University of Connecticut, USA, in 2014
and 2016, respectively. He is an assistant research
professor in electrical and computer engineering at
the University of Connecticut.
His research interests include mathematical opti-
mization, operations, and economics of electricity markets.
Yaowen Yu (S12) received his B.S. degree in
automation from Huazhong University of Science
and Technology, Wuhan, China, in 2011, and Ph.D.
degree from University of Connecticut in 2016. He
is currently a senior application engineer at ABB
Enterprise Software, San Jose, CA.
His research interests include power system opti-
mization, grid integration of renewable energy, and
economics of electricity markets.